- Unsupervised Learning
- Posts
- Unsupervised Learning NO. 472
Unsupervised Learning NO. 472
STANDARD EDITION: 28 Open Cyber Jobs, Real-world AI Propaganda Poisoning, MCP Explained, Cline vs. Windsurf, and more...

Hey! I hope your week’s going well,
Updates on this side…
I have an RSS feed (bring it back!). RSS
🫶🏼 Thank you for all the cooking advice. Around 500 people responded with experience, encouragement, and recommendations! I ended up getting both a pressure cooker and an air fryer, and am going through all your responses now for recipe ideas. Really appreciate you all.
I’ve tried Windsurf but didn’t like it much. Cline is my favorite AI coding assistant. I just wish its UI was as nice as Cursor’s. Cline, please make an IDE like Windsurf and Cursor did.
My personal AI Infrastructure is getting insane now. Going to have something to show soon! Hints: One main agent —> All my personal service tooling —> various output types. In other words, Life Augmentation with AI, like we’ve been talking about. Basically this —> WHERE IT’S ALL GOING
MY WORK
My concise explanation of what Model Context Protocol (MCP) Servers are, and why everyone’s so excited about them. THREAD
On this topic—a bit of a personal feel-good moment here: I wrote this in 2016:
This is essentially what MCPs are! This is finally starting to happen. MCPs a way to take any application and turn it into globally available services.
So all the pieces are:
We all get Digital Assistants that know everything about us
Everything gets an API, including people, objects, and businesses
Our DAs are the ones interacting with the world on our behalf, because there will be billions of APIs (Daemons) out there to interact with
Our DAs will display the current best contextual view in our AR interfaces
Well, everyone is currently building DAs. MCPs and lots of other projects are the APIs for everything. And lots of people are working on the AR stuff as well.
It’s happening! So excited for this future. Assuming we make it, lol.
Sponsor
AI is already in your organization—are you in control?
AI adoption is happening whether you have a plan or not.
Do you know what data is feeding AI models? Can you see how AI is being used across your business? Is your current data protection effective enough to prevent exposure?
Harmonic Security offers clear visibility into AI risks and gives you the controls you need to manage them—without slowing innovation.
👋🏼 I have had the opportunity to chat at length with the founder of Harmonic, and I love the way they’re addressing the AI data leakage issue. 👆🏼
CYBERSECURITY
A critical PHP-CGI remote code execution vulnerability (CVE-2024-5477) is being mass-exploited against Windows systems globally. It started with a bunch of Japanese targets and it’s now gone global. LINK
HelpNetSecurity put together a list of 28 (not fake) cybersecurity positions across the United States. They range from CISOs to Threat Intelligence interns and lots in-between. LINK
The President has nominated Sean Plankey, a veteran with experience at U.S. Cyber Command and the previous Trump administration, to head CISA. LINK
Malicious Poisoning of AI Narratives
TechCrunch reports that Russian propaganda network "Pravda" is flooding the web with fake news that has successfully poisoned AI responses from ChatGPT and other AI chatbots. LINK | AMERICAN SUNLIGHT PROJECT
- 10 leading chatbots were tested for susceptibility to Russian disinformation
- Pravda published 3.6 million misleading articles in 2024 alone
- "NewsGuard's analysis found that chatbots collectively repeated false Russian disinformation narratives 33% of the time."
💡 This is a fascinating real-world AI attack. It used to be about SEO poisoning, and now it’s about flooding the internet with marketing/propaganda that will get picked up by AI bots.
This is the most realistic and dangerous AI supply chain attack I’ve seen so far.
Turns out it’s not about attacking the initial training sets; it’s about attacking the media space which gets crawled by the AI bots and results in updates.
And it might be Russian attackers doing this particular thing, but you could argue that this has been done for years already with common narratives around things.
The point is whatever everyone is talking about, and sharing, and writing online—becomes what an internet-crawling AI thinks is real.
So narrative memes become real, according to AI. It’s not smart enough to detect if people aren’t thinking for themselves and are just parroting their tribe’s talking points.
—
Poland's space agency POLSA has taken its systems offline after detecting unauthorized network access in what their digital minister says is part of ongoing Russian cyber campaigns against the country. LINK
Japanese telecom giant NTT got breached and exposed corporate data for around 18,000 companies. LINK
AI Targeting of “Enemies”
The US is using AI to scan social media and revoke visas of foreign students it believes support Hamas or other terrorist groups. This might sound ok to a lot of people, but you have to remember the “first they came for” lesson. It starts with obvious terrorist supporters, if things go bad that might expand to lots of different types of political opponents. LINK
💡 The whole thing also gets more worrying the more powerful AI becomes. As I’ve been talking about, one way to view Agents/Tools is the ability to apply thousands or millions of new sets of eyes/brains/hands to any problem.
Including seeking out political enemies.
In 2022 (pre-AI) it would require an extraordinary number of people to investigate the millions of blogs and YouTube channels to see who was talking bad about the government. And some of those people might object to doing such investigations.
With AI in the next months and years it’ll be possible to hire hundreds of thousands of “investigators” who can 1) collect, 2) analyze, and 3) create recommendations on who to go after. And they won’t object to anything.
So the prospect of searching all public discourse to find and target political enemies is now becoming possible for the first time.
—
The DOJ charged 12 Chinese nationals for state-backed hacking operations including two government officers, eight employees of a hacking company called i-Soon, and a few members of the APT27 threat group. LINK
Microsoft Threat Intelligence says the China-linked Silk Typhoon hacking group is shifting tactics to compromise IT supply chains as an entry point to corporate networks. LINK
NATIONAL SECURITY
Russia and China (and others?) are targeting DOGE-fired US federal workers, with special focus on those with security clearances and possible knowledge of critical infrastructure. LINK
China announced another 7.2% increase to its military budget, taking it to $245 billion. LINK
Polish President Andrzej Duda is pushing NATO members to increase defense spending to 3% of GDP during his visit to Brussels. LINK
Oliver Carroll reports that the US has cut off crucial HIMARS intelligence to Ukraine, reducing their ability to defend against Russian forces. LINK
Sponsor
MTTC: The SOC Metric That Changes Everything
SOC speed isn’t just about detecting threats - it’s about how quickly you resolve them.
Enter MTTC (Mean Time to Conclusion), the AI-driven metric that reduces backlog, alert fatigue, and risk exposure.
If your SOC isn’t tracking MTTC, you’re already behind.
Get the full breakdown in this free guide.
AI
Manus Breaks AI Internet
The big hype news this last couple of weeks has been Manus, which is leaning pretty hard towards AGI. It’s basically like Claude or OpenAI or any of the other systems, but people are saying it’s just way smarter at following the plot and achieving crazy-good outcomes.
The product is very much in-line with a post I made the other day about AGI, and how I don’t think we need any smarter models, just better orchestration. And this is what a lot of people are saying too—basically this just feels like Claude underneath, but better.
I really want people to think deeply about this. I think GENERAL intelligence/competence might already be possible given how smart something like o1-pro
or Sonnet 3.7
already are. It already has the intelligence to understand millions (billions?) of general questions and know how it would proceed in solving them.
The problem is it not having enough context, or the right tools, and getting confused with too many threads. This isn’t an intelligence problem, per say, because it knew what it needed to do. It’s a coordination and complexity problem, which is more about working memory than anything else—or at least that’s my strong intuition.
Something to watch for sure. And this matters a lot because if I’m right, it means we’re very close to human-replacing AGI that will disrupt hundreds of millions of jobs globally.
Think of it this way—we can have moderately intelligent models (like we have today), and way better working memory and agent orchestration, or we can get way smarter models. But we probably don’t need both even though we’re definitely getting both.
But if one was more likely than the other, it’d be the orchestration at current intelligence levels.
—
The QwQ-32B
AI Model
Alibaba's new 32B parameter open-source reasoning model achieves performance comparable to DeepSeek's 671B parameter model thanks to reinforcement learning. LINK | GET IT ON OLLAMA
This is the most impressive local model I’ve ever used, and it’s not even close. Not sure why this isn’t being talked about more. It’s a thinking model and it’s just been extraordinary in my testing.
ollama run qwq
LINK
—
Mayo Addresses AI Hallucinations
Mayo Clinic is using "reverse RAG" to prevent hallucinations in their AI systems by linking every data point back to its source. LINK
- Reverse RAG is basically fact validation on AI claims
- "Physicians are very skeptical, and they want to make sure that they're not being fed information that isn't trustworthy. So trust for us means verification of anything that might be surfaced as content." - Matthew Callstrom
- Tasks that took doctors 90 minutes now completed in about 10 minutes
—
Tyler Cowen told Dwarkesh Patel that he predicts AI will only boost economic growth by 0.5% per year. Half a percent? I disagree massively, and so does Dwarkesh. LINK | YOUTUBE CONVO
Black Forest Labs (Replicate) is dominating image generation while DALL-E has dropped 80% in usage. Definitely go play with Replicate if you haven’t yet. My go-to is still Midjourney, though I am looking closely at Replicate now. LINK
Larry Page is working on a new AI company that uses artificial intelligence to design and manufacture optimized products. Basically start to finish design to manufacturing. LINK
Ben Buchanan, Biden's AI advisor, says Washington now believes AGI is possible and realizes they're competing with tech companies for talent to regulate it. LINK
Nirvana raised $80M for their AI-based trucking insurance platform that uses telematics and 20 billion miles of driving data to create better policies for drivers. This is the way forward for pretty much all products: More data, more context, more intelligence. Better product. But especially insurance. LINK | MY POST ON REALTIME INSURANCE PREMIUMS FROM 2016
Simon Willison taught a NICAR 2025 workshop on advanced web scraping techniques. Heavy focus on anti-scraping mitigation through screenshots, dom-loading, etc. LINK
McDonald's is rolling out a massive AI upgrade for its 43,000 locations, including AI drive-throughs, smart kitchen equipment, and a "generative AI virtual manager". This will be one of the biggest visible AI rollouts so far. Will be fascinating to see how it works. I unfortunately have experience with Dominos and Carls Jr.’s AI, and it’s quite good. LINK
TECHNOLOGY
Apple Delays AI Siri Upgrade
Apple is pushing back its "more personalized" Siri features because the development is "taking longer than we thought." Which basically means Siri having access to your full data context across your whole Apple ecosystem. LINK
A lot of people think this means no Apple Intelligence until then, but I’m already using tons of the new intelligence stuff and they’re great. Especially ChatGPT integration with Siri. So now I just ask Siri things and she can give me realtime data back. That’s huge by itself and it is already out, so don’t think it’s all delayed.
There’s also speculation that the delay is due to prompt injection concerns, which I think is probably directionally correct. I mean—what could go wrong with user-controlled querying of your entire life context? This is very security-sensitive for all the reasons we’ve been talking about here for years.
—
Waymo is expanding its self-driving car service into Palo Alto and some surrounding peninsula areas. So happy about this. Can’t wait til they’re in my town. I do feel bad for the Uber drivers, though. LINK
New Zealand appears to be managing its $16 billion health budget using a single Excel spreadsheet. Very DOGEy. LINK
Apple is also evidently planning a major redesign of iOS, iPadOS, and macOS for the Fall, redoing everything from icons to menus and system buttons. LINK
Tesla's global sales are tanking pretty hard right now, with significant drops across Germany, Australia, and China—not just the US. LINK | HACKER NEWS DISCUSSION
The new Mac Studio boasts M3 Ultra with 32 CPU cores and enough unified memory (512GB!) to load language models with 600 billion parameters. I got excited for a second, but I already have a dedicated AI machine (2 × 4090’s) and I think I’m going to sit out this cycle. LINK
HUMANS
Multiple indicators including Nowcast models and the re-inverted yield curve suggest we might face a higher-than-expected recession risk going into 2025. LINK
Brad Sigmon is set to become the first person executed by firing squad in the US since 2010 unless South Carolina's governor, or the Supreme Court, intervenes. LINK
Tiago Forte says traditional goal-setting may be killing our creativity and suggests optimizing for novelty and interestingness instead of fixed outcomes. LINK
Apple's hit dystopian show Severance is a great critique of how tech companies weaponize perks and culture to control employees. Having worked at lots of big companies, and been in management there, so much of this stuff is eerily accurate. LINK
NASA's new SPHEREx observatory is heading to space soon to map over 450 million galaxies. LINK
Stanford researchers have developed an antibody duo therapy that neutralizes all SARS-CoV-2 variants by targeting two different parts of the virus simultaneously. LINK
Researchers suggest autism and ADHD frequently coexist as "AuDHD". Studies show up to half of autistic people exhibit ADHD symptoms and two-thirds of people with ADHD show autism characteristics. LINK
Scientists at Mass Eye and Ear have used patients' own stem cells to repair cornea damage previously thought to be permanently blinding. LINK
Physicists have discovered a third limitation on our ability to predict the universe—in addition to quantum uncertainty and chaotic butterfly effects. It's called undecidability, and it means some systems are fundamentally unknowable. LINK
The price of coffee has gone up so much that the millions of businesses in its supply chain are running out of money to buy it. LINK
DISCOVERY
A FAANG machine learning engineer named Jasper Gilley quit his job because he believes AI will automate his entire role by the end of 2025. LINK
Robin Moffatt argues that we shouldn't just write the perfectly structured blog posts but also share our messy struggles and solutions that others might need too. LINK
Small docs
— A project around writing small docs instead of giant ones that are hard to follow. LINK
Jamie Rumbelow created a slick hybrid between databases, programming, and UIs that makes scripts more user-friendly without needing a full web app. LINK
AstroNvim
— One of the better Vim distros out there. I prefer LazyVim though, but hate the name. LINK
InterviewDB
— Lets you access crowdsourced tech interview questions from the community and sources like 1point3acres, so you don't have to hunt for them everywhere. LINK
Alexander Solzhenitsyn on the future of the West. LINK
Senechal-MCP
— An MCP by a brilliant UL Member that helps LLMs access standardized health data from the Senechal API. LINK
Stealth Interview
— An AI tool that runs invisibly during technical interviews, helping you with coding questions while your screen is shared, directly challenging the Leetcode gatekeeping system. Or helping people cheat, depending on perspective. LINK | HACKER NEWS DISCUSSION
Piano.fm
— Generate endless, beautiful AI piano music that adapts to your preferences and never repeats. LINK | HACKER NEWS DISCUSSION
Matrix Hacking Game
— Someone built a Matrix-themed AI hacking game that lets users learn about prompt engineering and security by defeating an Agent Smith AI character. LINK | HAKCER NEWS DISCUSSION
Sen
— A new app connects your iPhone camera to multimodal LLMs for real-time visual intelligence, similar to what you might do with ChatGPT but instant and mobile. Made with $0 VC money and Cursor. This is the way! LINK | HACKER NEWS DISCUSSION
Sense Feedback
— This dev built an AI that reads the emotional vibe of your product's social media mentions and gives you instant feedback. LINK | HACKER NEWS DISCUSSION
Someone’s pretty decent Cursor rules. LINK
Is Intelligence Just Interest?
What if people being “smart” or “brilliant” about something, really just comes down to being obsessed with it?
When I think about most people I know who are so-called geniuses about a given topic, it tends to overlap 100% with people who won’t shut up about it. Kind of a weird coincidence, no?
I’m thinking back to training people as well, in cybersecurity specifically, and it turned out that the people who ended up thriving were the ones who simply loved it, and were curious about it, which made them want to study more.
So is curiosity actually the superpower?
GET THIS KIND OF MEMBER CONTENT PLUS ALL OTHER MEMBER SECTIONS EVERY WEEK. UPGRADE NOW.
Upgrade to the Member Edition
You’re currently receiving the STANDARD edition. Subscribers to the MEMBER Edition get additional content, including IDEAS, a BI-MONTHLY ESSAY, the DISCOVERY section full of the best content I’ve seen this week, the RECOMMENDATION OF THE WEEK, and the APHORISM of the WEEK.
In addition, you get access to the UL Member Community, which includes private chat with 1000+ of the smartest and kindest members you’ll find anywhere on the internet.